Research On Shereo Matching Algorithm Based On SOM

Binocular stereo vision can be widely used in 3D measurement,machine navigation,human-computer interaction and other fields by imitating human perception of space.Stereo matching is an important part of 3D reconstruction.The development of high-precision stereo matching algorithms is the basis for achieving precise spatial reconstruction.The accuracy of current stereo matching algorithms is easily affected by factors such as occlusion,illumination,depth discontinuities,and weakly textured areas.Therefore,obtaining high-precision parallax maps is still very difficult.In view of the problem of low matching accuracy in weak texture areas and deep discontinuities,we have improved the local stereo matching algorithm,and proposed to introduce self-organizing map(Self-Organizing feature Map,SOM)neural network into stereo matching image segmentation and confidence propagation algorithms are combined to obtain high-precision stereo matching disparity maps.The main contents are as follows:(1)Aiming at the problem that the accuracy of the segmentation matching algorithm is affected by the segmentation effect,we propose to apply the SOM segmentation method to the matching algorithm to improve the accuracy of the matching result by obtaining accurate segmented images.The experiment compares the SOM segmentation method and other clustering-based segmentation methods,and compares the actual efficacy of the SOM segmentation method and other segmentation methods on the accuracy of stereo matching,which proves the feasibility and superiority of the algorithm in this paper.(2)Aiming at the problem of using a single similarity measure for matching cost to make the matching accuracy lower,a matching cost calculation method is proposed which combines multiple similarity measures.The algorithm in this paper improves the traditional Census matching method,combines three complementary similarity measures to establish a joint matching cost,and then obtains the initial disparity ruap through matching cost aggregation and disparity calculation.It is proved through experiments that the improved local matching algorithm improves the matching accuracy in the repeated texture area and weak texture area.(3)Use trusted pixel screening to classify regional pixels to obtain accurate parallax plane templates,and apply regional-based confidence propagation algorithms for global parallax optimization based on the acquired parallax plane templates.Comparison experiments with similar algorithms show that our algorithm has a good matching effect in the weak texture area and the depth discontinuity area.